An independent researcher named Derik Kauffman has discovered that Apple embeds a hard-coded "grey list" of more than 100,000 words and phrases into iOS that are deliberately excluded from predictive text and swipe-typing completion. The list, encoded directly in the operating system, means users must manually spell out anything on it — no autocomplete, no swipe shortcuts. The discovery came after Kauffman noticed pattern failures in his own typing and traced the behavior to a specific table in Apple's code. Some entries are predictable content-policy fare: drug names, sexual terms, slurs. Others are not. The list suppresses "Trayvon Martin," "Kashmir dispute," "Muslim extremist," "male chauvinist pig," and "Dictator Trump." In Hebrew, "the occupied territories" is blocked. In Arabic, "the Hamas movement" is blocked. The suppression extends to benign-sounding fragments: "Australia is a very bad place," "question your sexuality," and "on MDMA" all fail swipe-typing. Attempting "United Ireland" via swipe produces "United Iceland." The list also contains what looks like corporate self-protection. "iMessage isn't working" will not autocomplete on an iPhone — a detail that reads less like content moderation and more like reputation management baked into the keyboard layer. Apple's statement in response said the keyboard is designed to avoid "inappropriate or offensive predictive suggestions" and that users can always type words out manually and override suggestions. This sits within a longer history of Apple's text-input engineering. Early autocorrect relied on simple dictionary lookups in Unix. N-gram statistical models followed, producing the infamous "ducking" substitutions. Last year Apple shipped a transformer-based autocorrect system — closer to an LLM — which Reddit users blamed for a wave of bizarre spelling suggestions. The grey list predates all of this: it is a blunt, manually curated override that sits on top of whatever statistical model is running. The core tension is between user agency and platform control. Apple frames this as protecting users from offensive suggestions. But the list's scope goes well beyond profanity filtering. Suppressing the name of a shooting victim, a geopolitical dispute, or a phrase critical of a political figure is a different category of intervention — one that shapes what is easy to say and what requires extra effort. On a device where swipe-typing is the dominant input mode, friction is not neutral. Making a phrase harder to type is not the same as censoring it, but it is not nothing either. Kauffman, whose other work involves using AI to detect errors in scientific papers, described the list to Reuters as a "light-touch form of censorship." That framing is debatable — but what is not debatable is that a single company, without public review or disclosed criteria, decides which of the world's words get keyboard assistance and which do not. The list is not published. The criteria for inclusion are not explained. Users are not told it exists. The only reason anyone knows about it is that one researcher noticed his keyboard behaving strangely and went digging. The deeper question is whether infrastructure-level language shaping — invisible, undisclosed, applied globally across languages and geographies — should be treated as a product design choice or as something that warrants transparency standards. Apple controls the default keyboard on over a billion active iPhones. When that keyboard quietly makes certain phrases harder to produce, the aggregate effect on discourse is non-trivial, even if any individual user can always just type the letters out.
